Int J CARS (2011) 6 (Suppl 1):S305–S366
DOI 10.1007/s11548-011-0613-1
COMPUTER ASSISTED RADIOLOGY - 25TH INTERNATIONAL CONGRESS AND EXHIBITION
Quantification of periarticular demineralization in rheumatoid
arthritis by digital X-ray radiogrammetry (DXR) and peripheral
quantitative computed tomography (pQCT)
D.M. Renz1, A. Pfeil2, A. Hansch3, G. Wolf 2, J. Böttcher4
1
Charité University Medicine Berlin, Department of Radiology,
Berlin, Germany
2
Friedrich-Schiller-University Jena, Department of Internal Medicine
III, Jena, Germany
3
Friedrich-Schiller-University Jena, Institute of Diagnostic
and Interventional Radiology, Jena, Germany
4
SRH Waldklinikum Gera, Institute of Diagnostic and Interventional
Radiology, Gera, Germany
Keywords DXR Rheumatoid arthritis Peripheral quantitative
computed tomography Bone mineral density
Purpose
Rheumatoid arthritis as a chronic and inflammatory disease of the hand
is associated with a periarticular bone loss [1, 2]. Advances in established techniques and the implementation of new computer-aided
diagnosis (CAD) methods in the field of osteodensitometry have
widened the spectrum of diagnostic tools available for assessing bone
mineral density (BMD) in different body regions. The most established
techniques used for estimating BMD in the routine clinical setting
include quantitative computed tomography, quantitative ultrasound,
and dual-energy X-ray absorptiometry (DXA). Digital X-ray radiogrammetry (DXR) is a recently developed CAD tool for automatically
assessing cortical mineralization. The aim of this study was to evaluate
changes of bone mineral density using the radiogrammetrically based
densitometric technology (DXR) in patients suffering from rheumatoid arthritis and to compare this technique with findings of peripheral
quantitative computed tomography (pQCT).
Methods
Ninety patients with a verified RA, following the criteria of the
American College of Rheumatology [3], underwent a prospective
analysis of their BMD. Bone mineralization was assessed using DXR
and pQCT. Digital X-ray radiogrammetry (DXR, Pronosco X-Posure
SystemTM, Version 2.0; Sectra, Linköping, Sweden) was applied to
quantify bone mineral density (BMD in g/cm2) based on radiographs
of the hand in an anterior-posterior projection. The radiographs were
digitized by a scanner. CAD analysis was performed without user
interaction. DXR-BMD was measured on the three middle metacarpals. To locate the diaphysis of the metacarpals in the radiographs,
Pronosco X-posure used a well-established model-based algorithm
known as the active shape model (ASM). After each diaphysis had
been identified, the computer algorithm automatically defined regions
of interest (ROIs) for measurements at the narrowest parts of the
diaphysis of the index, middle, and ring finger metacarpals. PQCTBMD (Version 3.3; Stratec Medizintechnik GmbH, Pforzheim, Germany) was calculated of the distal radius and differentiated in total,
trabecular, and cortical BMD. The clinical severity of RA was
assessed using the Larsen Score.
Results
The mean value of DXR-BMD decreased from 0.57 g/cm2 ± 0.08
(Larsen Score 1) to 0.45 g/cm2 ± 0.11 (Larsen Score 5). The relative
decrease of BMD measured by DXR between the highest and the
lowest score was 20% (p \ 0.05). The relative decrease of BMD
(pQCT) from Larsen Score 1 to Score 5 also showed a significant
result regarding pQCT-BMD (trabecular; as the most metabolic active
bone tissue) with 16% (p \ 0.05). No significant changes in the
demineralization were confirmed for pQCT-BMD (total) with 12%
and for pQCT-BMD (cortical) with 2%.
Conclusion
The development of computer-based techniques has promoted the
precise quantification of metacarpal bone mineral density as measured
by digital X-ray radiogrammetry [4]. The study highlights that DXR
is able to exactly measure cortical differences of bone mineralization
in patients suffering from rheumatoid arthritis. This CAD technique
seems to be able to reliably quantify disease-related periarticular loss
of bone mineral density dependent on the severity of rheumatoid
arthritis. Furthermore, the consideration of DXR parameters seems to
be an outcome measure in RA and seems to allow an optimal planning
and monitoring of therapeutic strategies of the disease [5].
References
[1] Gravallese EM. Bone destruction in arthritis. Ann Rheum Dis
2002; 61: 84–86.
[2] Goldring SR. Periarticular bone changes in rheumatoid arthritis:
pathophysiological implications and clinical utility. Ann Rheum
Dis 2009; 68: 297–299.
[3] Arnett FC, Edworthy SM, Bloch DA, McShane DJ, Fies FJ,
Cooper NS. The American Rheumatism Association 1987
revised criteria for the classification of rheumatoid arthritis.
Arthritis Rheum 1988; 31: 315–324.
[4] Böttcher J, Pfeil A, Rosholm A, Petrovitch A, Seidl BE, Malich
A, Schäfer ML, Kramer A, Mentzel HJ, Lehmann G, Hein G,
Kaiser WA. Digital X-ray radiogrammetry combined with
semiautomated analysis of joint space widths as a new
diagnostic approach in rheumatoid arthritis: a cross-sectional
and longitudinal study. Arthritis Rheum 2005; 52: 3850–3859.
[5] Pfeil A, Haugeberg G, Hansch A, Renz DM, Lehmann G,
Malich A, Wolf G, Böttcher J. The value of digital X-ray
radiogrammetry in the assessment of inflammatory bone loss in
rheumatoid arthritis. Arthritis Care Res (Hoboken). 2011 Jan 4
[Epub ahead of print].
Respiratory motion prediction by evolving connectionist system
(ECOS) framework
M. Kakar
Oslo University Hospital, Division of Surgery and Cancer Medicine,
Institute for Cancer Research, Department of Radiation Biology,
Oslo, Norway
Keywords Respiratory Motion prediction Evolving EFuNN DENFIS
Purpose
Respiratory motion prediction is a chaotic time series prediction
problem. In this study, respiratory motion predictability from 10
123
S306
Int J CARS (2011) 6 (Suppl 1):S305–S366
traces from a lung cancer patient is examined by using two evolving
fuzzy neural architectures. The main objective has been to test
algorithms which can evolve with the breathing characteristics of the
patient for online implementation.
Methods
Previous studies [1] have shown that popular Neural network models
including Multilayer Percepron (MLP) and Adaptive Neuro Fuzzy
Inference System (ANFIS) are good for modeling and predicting
respiratory motion, however, they lack in on line learning and capability to evolve. These models operate on a fixed size connectionist
structure, which limits its ability to accommodate new data, and may
require much iteration for learning. In addition, it is required that the
initialization parameters of the network be tested with multiple initial
values. Such an approach is cumbersome and renders impractical
especially for online implementations. As an alternative, evolving
fuzzy neural architectures are becoming popular for online training
[2]. These require much larger dataset for modeling and emphasizes
on learning specific/local characteristics of the process rather than
global characteristics.
Breathing data curves were obtained from Real Time Position
Management system (RPM system) from 10 traces from a single
patient at different time intervals [3]. Two different methods in the
evolving connectionist’s system framework (ECOS) including
Dynamic Evolving neuro fuzzy inference system (DENFIS) and
Evolving fuzzy neural network (EFuNN) were tested on our dataset
for a 200 ms prediction interval.
EFuNN is a five layered fuzzy neural network. EFuNN evolves
according to ECOS principles. ECOS evolves structurally and functionally by interacting with the environment thereby better adapting to
the incoming data. The nodes and connections in the layers in EFuNN
are created or connected based upon the data samples being presented. In our simulations, we chose sensitivity threshold as 0.9 and
Table 1 Average RMSE values for learning and validation for different architectures (1 step ahead prediction i.e. 200 ms)
Average RMSE
DENFIS
EfuNN
Training
21.15
8.66
Testing
21.51
4.84
error threshold as 0.4. Number of membership functions was taken to
be 3. Learning rate for weights w1 and w2 was 0.1.
DENFIS uses Takagi–Sugeno type fuzzy inference engine[4]. The
inference engine here is composed of m rules. The learning mechanism
relies on dynamically creating and updating fuzzy rules during the
learning process. The fuzzy rules that participate in the inference for
each of the existing input vector are dynamically chosen from the
existing fuzzy rule set depending upon the position of the current input
vector in the input space. In this study, after experimenting with different
combinations, we found the best choices for Dthras 0.1 and number of
epochs for training as 2. The number of nodes which are referenced to
estimate the output of the current sample was chosen to be 3.
Results
EFuNN and DENFIS algorithms were tested on 10 traces of a
respiratory motion sequences. Interval of the signal was chosen to be
random in order to introduce generality for testing. Results obtained
from simulation over 10 traces of data are listed in table 1.Results
reveal root mean square error (RMSE) values to be 21.51 and 4.84 for
prediction by DENFIS and EFuNN, respectively. Figure 1 shows a
sample for prediction by EFuNN (continuous line) w.r.t. desired
(dashed) output for a single trace with 1558 breathing samples.
Conclusion
EFuNN was found to have better performance than DENFIS on this
dataset.
References
[1] Kakar M., Nystrom H., Aarup L.R., Nottrup T.J., and Olsen
D.R., ‘‘Respiratory motion prediction by using the adaptive
neuro fuzzy inference system (ANFIS),’’ Phys. Med. Biol., vol.
50, no. 19, pp. 4721–4728, 2005.
[2] Watts M.J., ‘‘A decade of Kasabov’s evolving connectionist
systems: A Review’’ IEEE Trans. on Systems, Man, and
Cybernetics, vol. 39, no. 3, pp. 253–269.
[3] Nøttrup T.J., Korreman S. S., Pedersen A. N., Aarup L. R.,
Nystrom H., Olsen M., and Specht L., ‘‘Intra- and interfraction
breathing variations during curative radiotherapy for lung
cancer,’’ Radiotherapy and Oncology, vol. 84, no. 1,
pp. 40–48, July2007.
[4] Takagi T. and Sugeno M., ‘‘Fuzzy identification of systems and
its application to modeling and control,’’ IEEE Trans. on
Systems, Man, and Cybernetics, no. 116, p. 132, 1985.
A framework for real-time target tracking in IGRT using threedimensional ultrasound
R. Bruder1, F. Ernst1, A. Schlaefer1, A. Schweikard1
1
University of Lübeck, Institute for Robotics and Cognitive Systems,
Lübeck, Germany
Keywords 3D ultrasound IGRT Radio surgery Tracking Template matching
Fig. 1 Desired (dashed line) and EFuNN predicted (continuous line)
for 1,558 breathing samples from a Lung patient
123
Purpose
Motion compensation is a well-known practice for radiation therapy
[1, 2]. As the target is usually not directly measurable or the sampling
rate is limited, surrogate signals are sampled with higher rates and
correlated to the internal target movement [3]. This method works
well for regular, periodic target motion and surrogate signals, which
reflect a high amount of this internal motion. For irregular or
changing motion patterns the results of this method are limited.
Especially in areas, where multiple motion sources overlap into one
target movement, e.g. in proximity of the heart or blood vessels,
correlation is nearly impossible or requires a high sampling rate of the
target region.
We concentrate on using three-dimensional ultrasound as tracking
modality for target localization. This method offers the possibility to